Non-Invasive Brain-Computer Interfaces: State of the Art and Trends

Abstract

Brain-computer interface (BCI) is a rapidly evolving technology that has the potential to widely influence research, clinical and recreational use. Non-invasive BCI approaches are particularly common as they can impact a large number of participants safely and at a relatively low cost. Where traditional non-invasive BCIs were used for simple computer cursor tasks, it is now increasingly common for these systems to control robotic devices for complex tasks that may be useful in daily life. In this review, we provide an overview of the general BCI framework as well as the various methods that can be used to record neural activity, extract signals of interest, and decode brain states. In this context, we summarize the current state-of-the-art of non-invasive BCI research, focusing on trends in both the application of BCIs for controlling external devices and algorithm development to optimize their use. We also discuss various open-source BCI toolboxes and software, and describe their impact on the field at large.

Fig. 1.

Neural recording modalities used for brain-computer interface applications. (top) Non-invasive techniques include magnetoencephalography (MEG), electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). (middle) Minimally invasive approaches involve electrocorticography (ECoG) and relatively novel technologies such as endovascular electrodes (i.e., the Stentrode) and functional ultrasound (fUS). (bottom) Invasive techniques include stereoEEG (sEEG) with penetrating electrodes and multi-unit arrays (i.e., the Utah Array). Nearly all these techniques measure electrophysiological signals with the exception of fNIRS and fUS, which measure an indirect hemodynamic readout of neuronal activity.

Fig. 2.

Overview of neural signals used for noninvasive brain-computer interface control. These signals can be broadly categorized according to being endogenous or exogenous in origin, frequency or time domain, and by the brain region/electrode coverage from which they are acquired. Motor imagery tasks generate an event-related (de)synchronization that can be detected from motor electrodes and which generate a velocity-based signal for end effector control direction. Overt spatial attention refers to a gaze-based action that is detected in electrodes covering the parietal cortex and which drives an end effector in a particular direction. Steady-state visual evoked potentials (SSVEPs) refer to increases in narrow band power in electrodes covering the visual cortices upon attending to a stimulus flickering at the corresponding frequency. The P300 response is elicited during an oddball-type context when a user’s choice is selected. Finally, slow cortical potentials refer to electrical potential deflections that are time-locked to movement events or that correspond to limb kinematics.

Fig. 3.

Neuro-structural model and general trial structure to elicit ERD/ERS. a) Neurostructural model for the description and interpretation of resulting EEG patterns. M1 and S1 symbolize primary motor and somatosensory cortex, respectively, and both carry bidirectional pathways to the deep brain and brainstem (DB/BS). M1 projects directly to the spinal cord, which is also connected to the DB/BS via ascending and descending fibers. Efferent and afferent connections in the peripheral nervous system are indicated by double lined arrows. b) The general structure of a task trial consists of a preparation phase (grey box), followed by the “Task” (red box) and a post-movement phase (blue box). Movement onset is indicated by the vertical black line.

Fig. 4.

BCI strategies leverage principles of the Yerkes-Dodson law to improve performance. Virtual and augmented reality paradigms have been integrated into BCI control to increase user engagement. By contrast, various mediation-based practices help reduce anxiety levels. In both cases, the user’s mental state is driven towards an optimal performance zone where skill acquisition and execution can be maximized.

Fig. 5.

EEG-BCI based control of a robotic arm using motor imagery tasks. Motor imagery was performed to control the “reach” of a robotic arm in one and two dimensions. Similar motor imagery protocols were also used to perform the “grasping” of a lego block. Task complexity increases from 2 to 5 during subject training. Reproduced from Fig. 1 of [163].

Fig. 6.

Overview of recent decoding methods used in EEG-based BCI. These methods can be broadly separated into five main categories: Deep learning, transfer learning, manifold classificiation, adaptive learning, and EEG source analysis. Each of these categories consists of detailed algorithms that aim to improve various aspects of EEG decoding and BCI performance. In the end, all of these approaches aim to develop robust, calibration-free and user-friendly BCIs.